1 /*
2 * Copyright 2012 Vadim Girlin <[email protected]>
3 * Authors:
4 * Vadim Girlin
5 * SPDX-License-Identifier: MIT
6 */
7
8 #ifndef R600_ISA_H_
9 #define R600_ISA_H_
10
11 #include "amd_family.h"
12 #include "util/u_debug.h"
13
14 #ifdef __cplusplus
15 extern "C" {
16 #endif
17
18 /* ALU flags */
19 enum alu_op_flags
20 {
21 AF_NONE = 0,
22 AF_V = (1<<0), /* allowed in vector slots */
23
24 /* allowed in scalar(trans) slot (slots xyz on cayman, may be replicated
25 * to w) */
26 AF_S = (1<<1),
27
28 AF_4SLOT = (1<<2), /* uses four vector slots (e.g. DOT4) */
29 AF_4V = (AF_V | AF_4SLOT),
30 AF_VS = (AF_V | AF_S), /* allowed in any slot */
31
32 AF_2SLOT = (1 << 3),
33 AF_2V = AF_V | AF_2SLOT, /* XY or ZW */
34
35 AF_KILL = (1<<4),
36 AF_PRED = (1<<5),
37 AF_SET = (1<<6),
38
39 /* e.g. MUL_PREV instructions, allowed in x/y, depends on z/w */
40 AF_PREV_INTERLEAVE = (1<<7),
41
42 AF_MOVA = (1<<8), /* all MOVA instructions */
43
44 AF_IEEE = (1<<10),
45
46 AF_DST_TYPE_MASK = (3<<11),
47 AF_FLOAT_DST = 0,
48 AF_INT_DST = (1<<11),
49 AF_UINT_DST = (3<<11),
50
51 /* DP instructions, 2-slot pairs */
52 AF_64 = (1<<13),
53 /* 24 bit instructions */
54 AF_24 = (1<<14),
55 /* DX10 variants */
56 AF_DX10 = (1<<15),
57
58 /* result is replicated to all channels (only if AF_4V is also set -
59 * for special handling of MULLO_INT on CM) */
60 AF_REPL = (1<<16),
61
62 /* interpolation instructions */
63 AF_INTERP = (1<<17),
64
65 /* LDS instructions */
66 AF_LDS = (1<<20),
67
68 /* e.g. DOT - depends on the next slot in the same group (x<=y/y<=z/z<=w) */
69 AF_PREV_NEXT = (1<<21),
70
71 /* int<->flt conversions */
72 AF_CVT = (1<<22),
73
74 /* commutative operation on src0 and src1 ( a op b = b op a),
75 * includes MULADDs (considering the MUL part on src0 and src1 only) */
76 AF_M_COMM = (1 << 23),
77
78 /* associative operation ((a op b) op c) == (a op (b op c)),
79 * includes MULADDs (considering the MUL part on src0 and src1 only) */
80 AF_M_ASSOC = (1 << 24),
81
82 AF_PRED_PUSH = (1 << 25),
83
84 AF_ANY_PRED = (AF_PRED | AF_PRED_PUSH),
85
86 AF_CMOV = (1 << 26),
87
88 // for SETcc, PREDSETcc, ... - type of comparison
89 AF_CMP_TYPE_MASK = (3 << 27),
90 AF_FLOAT_CMP = 0,
91 AF_INT_CMP = (1 << 27),
92 AF_UINT_CMP = (3 << 27),
93
94 /* condition codes - 3 bits */
95 AF_CC_SHIFT = 29,
96
97 AF_CC_MASK = (7U << AF_CC_SHIFT),
98 AF_CC_E = (0U << AF_CC_SHIFT),
99 AF_CC_GT = (1U << AF_CC_SHIFT),
100 AF_CC_GE = (2U << AF_CC_SHIFT),
101 AF_CC_NE = (3U << AF_CC_SHIFT),
102 AF_CC_LT = (4U << AF_CC_SHIFT),
103 AF_CC_LE = (5U << AF_CC_SHIFT),
104 };
105
106 /* flags for FETCH instructions (TEX/VTX/GDS) */
107 enum fetch_op_flags
108 {
109 FF_GDS = (1<<0),
110 FF_TEX = (1<<1),
111
112 FF_SETGRAD = (1<<2),
113 FF_GETGRAD = (1<<3),
114 FF_USEGRAD = (1<<4),
115
116 FF_VTX = (1<<5),
117 FF_MEM = (1<<6),
118
119 FF_SET_TEXTURE_OFFSETS = (1<<7),
120 FF_USE_TEXTURE_OFFSETS = (1<<8),
121 };
122
123 /* flags for CF instructions */
124 enum cf_op_flags
125 {
126 CF_CLAUSE = (1<<0), /* execute clause (alu/fetch ...) */
127 CF_ACK = (1<<1), /* acked versions of some instructions */
128 CF_ALU = (1<<2), /* alu clause execution */
129 CF_ALU_EXT = (1<<3), /* ALU_EXTENDED */
130 CF_EXP = (1<<4), /* export (CF_ALLOC_EXPORT_WORD1_SWIZ) */
131 CF_BRANCH = (1<<5), /* branch instructions */
132 CF_LOOP = (1<<6), /* loop instructions */
133 CF_CALL = (1<<7), /* call instructions */
134 CF_MEM = (1<<8), /* export_mem (CF_ALLOC_EXPORT_WORD1_BUF) */
135 CF_FETCH = (1<<9), /* fetch clause */
136
137 CF_UNCOND = (1<<10), /* COND = ACTIVE required */
138 CF_EMIT = (1<<11),
139 CF_STRM = (1<<12), /* MEM_STREAM* */
140
141 CF_RAT = (1<<13), /* MEM_RAT* */
142
143 CF_LOOP_START = (1<<14)
144 };
145
146 /* ALU instruction info */
147 struct alu_op_info
148 {
149 /* instruction name */
150 const char *name;
151 /* number of source operands */
152 int src_count;
153 /* opcodes, [0] - for r6xx/r7xx, [1] - for evergreen/cayman
154 * (-1) if instruction doesn't exist (more precise info in "slots") */
155 int opcode[2];
156 /* slots for r6xx, r7xx, evergreen, cayman
157 * (0 if instruction doesn't exist for gfx level) */
158 int slots[4];
159 /* flags (mostly autogenerated from instruction name) */
160 unsigned int flags;
161 };
162
163 /* FETCH instruction info */
164 struct fetch_op_info
165 {
166 const char * name;
167 /* for every gfx level */
168 int opcode[4];
169 int flags;
170 };
171
172 /* CF instruction info */
173 struct cf_op_info
174 {
175 const char * name;
176 /* for every gfx level */
177 int opcode[4];
178 int flags;
179 };
180
181
182 #define ALU_OP2_ADD 0
183 #define ALU_OP2_MUL 1
184 #define ALU_OP2_MUL_IEEE 2
185 #define ALU_OP2_MAX 3
186 #define ALU_OP2_MIN 4
187 #define ALU_OP2_MAX_DX10 5
188 #define ALU_OP2_MIN_DX10 6
189 #define ALU_OP2_SETE 7
190 #define ALU_OP2_SETGT 8
191 #define ALU_OP2_SETGE 9
192 #define ALU_OP2_SETNE 10
193 #define ALU_OP2_SETE_DX10 11
194 #define ALU_OP2_SETGT_DX10 12
195 #define ALU_OP2_SETGE_DX10 13
196 #define ALU_OP2_SETNE_DX10 14
197 #define ALU_OP1_FRACT 15
198 #define ALU_OP1_TRUNC 16
199 #define ALU_OP1_CEIL 17
200 #define ALU_OP1_RNDNE 18
201 #define ALU_OP1_FLOOR 19
202 #define ALU_OP2_ASHR_INT 20
203 #define ALU_OP2_LSHR_INT 21
204 #define ALU_OP2_LSHL_INT 22
205 #define ALU_OP1_MOV 23
206 #define ALU_OP0_NOP 24
207 #define ALU_OP2_PRED_SETGT_UINT 25
208 #define ALU_OP2_PRED_SETGE_UINT 26
209 #define ALU_OP2_PRED_SETE 27
210 #define ALU_OP2_PRED_SETGT 28
211 #define ALU_OP2_PRED_SETGE 29
212 #define ALU_OP2_PRED_SETNE 30
213 #define ALU_OP1_PRED_SET_INV 31
214 #define ALU_OP2_PRED_SET_POP 32
215 #define ALU_OP0_PRED_SET_CLR 33
216 #define ALU_OP1_PRED_SET_RESTORE 34
217 #define ALU_OP2_PRED_SETE_PUSH 35
218 #define ALU_OP2_PRED_SETGT_PUSH 36
219 #define ALU_OP2_PRED_SETGE_PUSH 37
220 #define ALU_OP2_PRED_SETNE_PUSH 38
221 #define ALU_OP2_KILLE 39
222 #define ALU_OP2_KILLGT 40
223 #define ALU_OP2_KILLGE 41
224 #define ALU_OP2_KILLNE 42
225 #define ALU_OP2_AND_INT 43
226 #define ALU_OP2_OR_INT 44
227 #define ALU_OP2_XOR_INT 45
228 #define ALU_OP1_NOT_INT 46
229 #define ALU_OP2_ADD_INT 47
230 #define ALU_OP2_SUB_INT 48
231 #define ALU_OP2_MAX_INT 49
232 #define ALU_OP2_MIN_INT 50
233 #define ALU_OP2_MAX_UINT 51
234 #define ALU_OP2_MIN_UINT 52
235 #define ALU_OP2_SETE_INT 53
236 #define ALU_OP2_SETGT_INT 54
237 #define ALU_OP2_SETGE_INT 55
238 #define ALU_OP2_SETNE_INT 56
239 #define ALU_OP2_SETGT_UINT 57
240 #define ALU_OP2_SETGE_UINT 58
241 #define ALU_OP2_KILLGT_UINT 59
242 #define ALU_OP2_KILLGE_UINT 60
243 #define ALU_OP2_PRED_SETE_INT 61
244 #define ALU_OP2_PRED_SETGT_INT 62
245 #define ALU_OP2_PRED_SETGE_INT 63
246 #define ALU_OP2_PRED_SETNE_INT 64
247 #define ALU_OP2_KILLE_INT 65
248 #define ALU_OP2_KILLGT_INT 66
249 #define ALU_OP2_KILLGE_INT 67
250 #define ALU_OP2_KILLNE_INT 68
251 #define ALU_OP2_PRED_SETE_PUSH_INT 69
252 #define ALU_OP2_PRED_SETGT_PUSH_INT 70
253 #define ALU_OP2_PRED_SETGE_PUSH_INT 71
254 #define ALU_OP2_PRED_SETNE_PUSH_INT 72
255 #define ALU_OP2_PRED_SETLT_PUSH_INT 73
256 #define ALU_OP2_PRED_SETLE_PUSH_INT 74
257 #define ALU_OP1_FLT_TO_INT 75
258 #define ALU_OP1_BFREV_INT 76
259 #define ALU_OP2_ADDC_UINT 77
260 #define ALU_OP2_SUBB_UINT 78
261 #define ALU_OP0_GROUP_BARRIER 79
262 #define ALU_OP0_GROUP_SEQ_BEGIN 80
263 #define ALU_OP0_GROUP_SEQ_END 81
264 #define ALU_OP2_SET_MODE 82
265 #define ALU_OP0_SET_CF_IDX0 83
266 #define ALU_OP0_SET_CF_IDX1 84
267 #define ALU_OP2_SET_LDS_SIZE 85
268 #define ALU_OP2_MUL_INT24 86
269 #define ALU_OP2_MULHI_INT24 87
270 #define ALU_OP1_FLT_TO_INT_TRUNC 88
271 #define ALU_OP1_EXP_IEEE 89
272 #define ALU_OP1_LOG_CLAMPED 90
273 #define ALU_OP1_LOG_IEEE 91
274 #define ALU_OP1_RECIP_CLAMPED 92
275 #define ALU_OP1_RECIP_FF 93
276 #define ALU_OP1_RECIP_IEEE 94
277 #define ALU_OP1_RECIPSQRT_CLAMPED 95
278 #define ALU_OP1_RECIPSQRT_FF 96
279 #define ALU_OP1_RECIPSQRT_IEEE 97
280 #define ALU_OP1_SQRT_IEEE 98
281 #define ALU_OP1_SIN 99
282 #define ALU_OP1_COS 100
283 #define ALU_OP2_MULLO_INT 101
284 #define ALU_OP2_MULHI_INT 102
285 #define ALU_OP2_MULLO_UINT 103
286 #define ALU_OP2_MULHI_UINT 104
287 #define ALU_OP1_RECIP_INT 105
288 #define ALU_OP1_RECIP_UINT 106
289 #define ALU_OP2_RECIP_64 107
290 #define ALU_OP2_RECIP_CLAMPED_64 108
291 #define ALU_OP2_RECIPSQRT_64 109
292 #define ALU_OP2_RECIPSQRT_CLAMPED_64 110
293 #define ALU_OP2_SQRT_64 111
294 #define ALU_OP1_FLT_TO_UINT 112
295 #define ALU_OP1_INT_TO_FLT 113
296 #define ALU_OP1_UINT_TO_FLT 114
297 #define ALU_OP2_BFM_INT 115
298 #define ALU_OP1_FLT32_TO_FLT16 116
299 #define ALU_OP1_FLT16_TO_FLT32 117
300 #define ALU_OP1_UBYTE0_FLT 118
301 #define ALU_OP1_UBYTE1_FLT 119
302 #define ALU_OP1_UBYTE2_FLT 120
303 #define ALU_OP1_UBYTE3_FLT 121
304 #define ALU_OP1_BCNT_INT 122
305 #define ALU_OP1_FFBH_UINT 123
306 #define ALU_OP1_FFBL_INT 124
307 #define ALU_OP1_FFBH_INT 125
308 #define ALU_OP1_FLT_TO_UINT4 126
309 #define ALU_OP2_DOT_IEEE 127
310 #define ALU_OP1_FLT_TO_INT_RPI 128
311 #define ALU_OP1_FLT_TO_INT_FLOOR 129
312 #define ALU_OP2_MULHI_UINT24 130
313 #define ALU_OP1_MBCNT_32HI_INT 131
314 #define ALU_OP1_OFFSET_TO_FLT 132
315 #define ALU_OP2_MUL_UINT24 133
316 #define ALU_OP1_BCNT_ACCUM_PREV_INT 134
317 #define ALU_OP1_MBCNT_32LO_ACCUM_PREV_INT 135
318 #define ALU_OP2_SETE_64 136
319 #define ALU_OP2_SETNE_64 137
320 #define ALU_OP2_SETGT_64 138
321 #define ALU_OP2_SETGE_64 139
322 #define ALU_OP2_MIN_64 140
323 #define ALU_OP2_MAX_64 141
324 #define ALU_OP2_DOT4 142
325 #define ALU_OP2_DOT4_IEEE 143
326 #define ALU_OP2_CUBE 144
327 #define ALU_OP1_MAX4 145
328 #define ALU_OP1_FREXP_64 146
329 #define ALU_OP2_LDEXP_64 147
330 #define ALU_OP1_FRACT_64 148
331 #define ALU_OP2_PRED_SETGT_64 149
332 #define ALU_OP2_PRED_SETE_64 150
333 #define ALU_OP2_PRED_SETGE_64 151
334 #define ALU_OP2_MUL_64 152
335 #define ALU_OP2_ADD_64 153
336 #define ALU_OP1_MOVA_INT 154
337 #define ALU_OP1_FLT64_TO_FLT32 155
338 #define ALU_OP1_FLT32_TO_FLT64 156
339 #define ALU_OP2_SAD_ACCUM_PREV_UINT 157
340 #define ALU_OP2_DOT 158
341 #define ALU_OP1_MUL_PREV 159
342 #define ALU_OP1_MUL_IEEE_PREV 160
343 #define ALU_OP1_ADD_PREV 161
344 #define ALU_OP2_MULADD_PREV 162
345 #define ALU_OP2_MULADD_IEEE_PREV 163
346 #define ALU_OP2_INTERP_XY 164
347 #define ALU_OP2_INTERP_ZW 165
348 #define ALU_OP2_INTERP_X 166
349 #define ALU_OP2_INTERP_Z 167
350 #define ALU_OP1_STORE_FLAGS 168
351 #define ALU_OP1_LOAD_STORE_FLAGS 169
352 #define ALU_OP2_LDS_1A 170
353 #define ALU_OP2_LDS_1A1D 171
354 #define ALU_OP2_LDS_2A 172
355 #define ALU_OP1_INTERP_LOAD_P0 173
356 #define ALU_OP1_INTERP_LOAD_P10 174
357 #define ALU_OP1_INTERP_LOAD_P20 175
358 #define ALU_OP3_BFE_UINT 176
359 #define ALU_OP3_BFE_INT 177
360 #define ALU_OP3_BFI_INT 178
361 #define ALU_OP3_FMA 179
362 #define ALU_OP3_MULADD_INT24 180
363 #define ALU_OP3_CNDNE_64 181
364 #define ALU_OP3_FMA_64 182
365 #define ALU_OP3_LERP_UINT 183
366 #define ALU_OP3_BIT_ALIGN_INT 184
367 #define ALU_OP3_BYTE_ALIGN_INT 185
368 #define ALU_OP3_SAD_ACCUM_UINT 186
369 #define ALU_OP3_SAD_ACCUM_HI_UINT 187
370 #define ALU_OP3_MULADD_UINT24 188
371 #define ALU_OP3_LDS_IDX_OP 189
372 #define ALU_OP3_MULADD 190
373 #define ALU_OP3_MULADD_M2 191
374 #define ALU_OP3_MULADD_M4 192
375 #define ALU_OP3_MULADD_D2 193
376 #define ALU_OP3_MULADD_IEEE 194
377 #define ALU_OP3_CNDE 195
378 #define ALU_OP3_CNDGT 196
379 #define ALU_OP3_CNDGE 197
380 #define ALU_OP3_CNDE_INT 198
381 #define ALU_OP3_CNDGT_INT 199
382 #define ALU_OP3_CNDGE_INT 200
383 #define ALU_OP3_MUL_LIT 201
384 #define ALU_OP1_MOVA 202
385 #define ALU_OP1_MOVA_FLOOR 203
386 #define ALU_OP1_MOVA_GPR_INT 204
387 #define ALU_OP3_MULADD_64 205
388 #define ALU_OP3_MULADD_64_M2 206
389 #define ALU_OP3_MULADD_64_M4 207
390 #define ALU_OP3_MULADD_64_D2 208
391 #define ALU_OP3_MUL_LIT_M2 209
392 #define ALU_OP3_MUL_LIT_M4 210
393 #define ALU_OP3_MUL_LIT_D2 211
394 #define ALU_OP3_MULADD_IEEE_M2 212
395 #define ALU_OP3_MULADD_IEEE_M4 213
396 #define ALU_OP3_MULADD_IEEE_D2 214
397
398 #define LDS_OP2_LDS_ADD 215
399 #define LDS_OP2_LDS_SUB 216
400 #define LDS_OP2_LDS_RSUB 217
401 #define LDS_OP2_LDS_INC 218
402 #define LDS_OP2_LDS_DEC 219
403 #define LDS_OP2_LDS_MIN_INT 220
404 #define LDS_OP2_LDS_MAX_INT 221
405 #define LDS_OP2_LDS_MIN_UINT 222
406 #define LDS_OP2_LDS_MAX_UINT 223
407 #define LDS_OP2_LDS_AND 224
408 #define LDS_OP2_LDS_OR 225
409 #define LDS_OP2_LDS_XOR 226
410 #define LDS_OP3_LDS_MSKOR 227
411 #define LDS_OP2_LDS_WRITE 228
412 #define LDS_OP3_LDS_WRITE_REL 229
413 #define LDS_OP3_LDS_WRITE2 230
414 #define LDS_OP3_LDS_CMP_STORE 231
415 #define LDS_OP3_LDS_CMP_STORE_SPF 232
416 #define LDS_OP2_LDS_BYTE_WRITE 233
417 #define LDS_OP2_LDS_SHORT_WRITE 234
418 #define LDS_OP2_LDS_ADD_RET 235
419 #define LDS_OP2_LDS_SUB_RET 236
420 #define LDS_OP2_LDS_RSUB_RET 237
421 #define LDS_OP2_LDS_INC_RET 238
422 #define LDS_OP2_LDS_DEC_RET 239
423 #define LDS_OP2_LDS_MIN_INT_RET 240
424 #define LDS_OP2_LDS_MAX_INT_RET 241
425 #define LDS_OP2_LDS_MIN_UINT_RET 242
426 #define LDS_OP2_LDS_MAX_UINT_RET 243
427 #define LDS_OP2_LDS_AND_RET 244
428 #define LDS_OP2_LDS_OR_RET 245
429 #define LDS_OP2_LDS_XOR_RET 246
430 #define LDS_OP3_LDS_MSKOR_RET 247
431 #define LDS_OP2_LDS_XCHG_RET 248
432 #define LDS_OP3_LDS_XCHG_REL_RET 249
433 #define LDS_OP3_LDS_XCHG2_RET 250
434 #define LDS_OP3_LDS_CMP_XCHG_RET 251
435 #define LDS_OP3_LDS_CMP_XCHG_SPF_RET 252
436 #define LDS_OP1_LDS_READ_RET 253
437 #define LDS_OP1_LDS_READ_REL_RET 254
438 #define LDS_OP2_LDS_READ2_RET 255
439 #define LDS_OP3_LDS_READWRITE_RET 256
440 #define LDS_OP1_LDS_BYTE_READ_RET 257
441 #define LDS_OP1_LDS_UBYTE_READ_RET 258
442 #define LDS_OP1_LDS_SHORT_READ_RET 259
443 #define LDS_OP1_LDS_USHORT_READ_RET 260
444
445 #define FETCH_OP_VFETCH 0
446 #define FETCH_OP_SEMFETCH 1
447 #define FETCH_OP_READ_SCRATCH 2
448 #define FETCH_OP_READ_REDUCT 3
449 #define FETCH_OP_READ_MEM 4
450 #define FETCH_OP_DS_LOCAL_WRITE 5
451 #define FETCH_OP_DS_LOCAL_READ 6
452 #define FETCH_OP_GDS_ADD 7
453 #define FETCH_OP_GDS_SUB 8
454 #define FETCH_OP_GDS_RSUB 9
455 #define FETCH_OP_GDS_INC 10
456 #define FETCH_OP_GDS_DEC 11
457 #define FETCH_OP_GDS_MIN_INT 12
458 #define FETCH_OP_GDS_MAX_INT 13
459 #define FETCH_OP_GDS_MIN_UINT 14
460 #define FETCH_OP_GDS_MAX_UINT 15
461 #define FETCH_OP_GDS_AND 16
462 #define FETCH_OP_GDS_OR 17
463 #define FETCH_OP_GDS_XOR 18
464 #define FETCH_OP_GDS_MSKOR 19
465 #define FETCH_OP_GDS_WRITE 20
466 #define FETCH_OP_GDS_WRITE_REL 21
467 #define FETCH_OP_GDS_WRITE2 22
468 #define FETCH_OP_GDS_CMP_STORE 23
469 #define FETCH_OP_GDS_CMP_STORE_SPF 24
470 #define FETCH_OP_GDS_BYTE_WRITE 25
471 #define FETCH_OP_GDS_SHORT_WRITE 26
472 #define FETCH_OP_GDS_ADD_RET 27
473 #define FETCH_OP_GDS_SUB_RET 28
474 #define FETCH_OP_GDS_RSUB_RET 29
475 #define FETCH_OP_GDS_INC_RET 30
476 #define FETCH_OP_GDS_DEC_RET 31
477 #define FETCH_OP_GDS_MIN_INT_RET 32
478 #define FETCH_OP_GDS_MAX_INT_RET 33
479 #define FETCH_OP_GDS_MIN_UINT_RET 34
480 #define FETCH_OP_GDS_MAX_UINT_RET 35
481 #define FETCH_OP_GDS_AND_RET 36
482 #define FETCH_OP_GDS_OR_RET 37
483 #define FETCH_OP_GDS_XOR_RET 38
484 #define FETCH_OP_GDS_MSKOR_RET 39
485 #define FETCH_OP_GDS_XCHG_RET 40
486 #define FETCH_OP_GDS_XCHG_REL_RET 41
487 #define FETCH_OP_GDS_XCHG2_RET 42
488 #define FETCH_OP_GDS_CMP_XCHG_RET 43
489 #define FETCH_OP_GDS_CMP_XCHG_SPF_RET 44
490 #define FETCH_OP_GDS_READ_RET 45
491 #define FETCH_OP_GDS_READ_REL_RET 46
492 #define FETCH_OP_GDS_READ2_RET 47
493 #define FETCH_OP_GDS_READWRITE_RET 48
494 #define FETCH_OP_GDS_BYTE_READ_RET 49
495 #define FETCH_OP_GDS_UBYTE_READ_RET 50
496 #define FETCH_OP_GDS_SHORT_READ_RET 51
497 #define FETCH_OP_GDS_USHORT_READ_RET 52
498 #define FETCH_OP_GDS_ATOMIC_ORDERED_ALLOC 53
499 #define FETCH_OP_TF_WRITE 54
500 #define FETCH_OP_DS_GLOBAL_WRITE 55
501 #define FETCH_OP_DS_GLOBAL_READ 56
502 #define FETCH_OP_LD 57
503 #define FETCH_OP_LDFPTR 58
504 #define FETCH_OP_GET_TEXTURE_RESINFO 59
505 #define FETCH_OP_GET_NUMBER_OF_SAMPLES 60
506 #define FETCH_OP_GET_LOD 61
507 #define FETCH_OP_GET_GRADIENTS_H 62
508 #define FETCH_OP_GET_GRADIENTS_V 63
509 #define FETCH_OP_GET_GRADIENTS_H_FINE 64
510 #define FETCH_OP_GET_GRADIENTS_V_FINE 65
511 #define FETCH_OP_GET_LERP 66
512 #define FETCH_OP_SET_TEXTURE_OFFSETS 67
513 #define FETCH_OP_KEEP_GRADIENTS 68
514 #define FETCH_OP_SET_GRADIENTS_H 69
515 #define FETCH_OP_SET_GRADIENTS_V 70
516 #define FETCH_OP_SET_GRADIENTS_H_COARSE 71
517 #define FETCH_OP_SET_GRADIENTS_V_COARSE 72
518 #define FETCH_OP_SET_GRADIENTS_H_PACKED_FINE 73
519 #define FETCH_OP_SET_GRADIENTS_V_PACKED_FINE 74
520 #define FETCH_OP_SET_GRADIENTS_H_PACKED_COARSE 75
521 #define FETCH_OP_SET_GRADIENTS_V_PACKED_COARSE 76
522 #define FETCH_OP_PASS 77
523 #define FETCH_OP_PASS1 78
524 #define FETCH_OP_PASS2 79
525 #define FETCH_OP_PASS3 80
526 #define FETCH_OP_SET_CUBEMAP_INDEX 81
527 #define FETCH_OP_GET_BUFFER_RESINFO 82
528 #define FETCH_OP_FETCH4 83
529 #define FETCH_OP_SAMPLE 84
530 #define FETCH_OP_SAMPLE_L 85
531 #define FETCH_OP_SAMPLE_LB 86
532 #define FETCH_OP_SAMPLE_LZ 87
533 #define FETCH_OP_SAMPLE_G 88
534 #define FETCH_OP_SAMPLE_G_L 89
535 #define FETCH_OP_GATHER4 90
536 #define FETCH_OP_SAMPLE_G_LB 91
537 #define FETCH_OP_SAMPLE_G_LZ 92
538 #define FETCH_OP_GATHER4_O 93
539 #define FETCH_OP_SAMPLE_C 94
540 #define FETCH_OP_SAMPLE_C_L 95
541 #define FETCH_OP_SAMPLE_C_LB 96
542 #define FETCH_OP_SAMPLE_C_LZ 97
543 #define FETCH_OP_SAMPLE_C_G 98
544 #define FETCH_OP_SAMPLE_C_G_L 99
545 #define FETCH_OP_GATHER4_C 100
546 #define FETCH_OP_SAMPLE_C_G_LB 101
547 #define FETCH_OP_SAMPLE_C_G_LZ 102
548 #define FETCH_OP_GATHER4_C_O 103
549
550 #define CF_OP_NOP 0
551 #define CF_OP_TEX 1
552 #define CF_OP_VTX 2
553 #define CF_OP_VTX_TC 3
554 #define CF_OP_GDS 4
555 #define CF_OP_LOOP_START 5
556 #define CF_OP_LOOP_END 6
557 #define CF_OP_LOOP_START_DX10 7
558 #define CF_OP_LOOP_START_NO_AL 8
559 #define CF_OP_LOOP_CONTINUE 9
560 #define CF_OP_LOOP_BREAK 10
561 #define CF_OP_JUMP 11
562 #define CF_OP_PUSH 12
563 #define CF_OP_PUSH_ELSE 13
564 #define CF_OP_ELSE 14
565 #define CF_OP_POP 15
566 #define CF_OP_POP_JUMP 16
567 #define CF_OP_POP_PUSH 17
568 #define CF_OP_POP_PUSH_ELSE 18
569 #define CF_OP_CALL 19
570 #define CF_OP_CALL_FS 20
571 #define CF_OP_RET 21
572 #define CF_OP_EMIT_VERTEX 22
573 #define CF_OP_EMIT_CUT_VERTEX 23
574 #define CF_OP_CUT_VERTEX 24
575 #define CF_OP_KILL 25
576 #define CF_OP_END_PROGRAM 26
577 #define CF_OP_WAIT_ACK 27
578 #define CF_OP_TEX_ACK 28
579 #define CF_OP_VTX_ACK 29
580 #define CF_OP_VTX_TC_ACK 30
581 #define CF_OP_JUMPTABLE 31
582 #define CF_OP_WAVE_SYNC 32
583 #define CF_OP_HALT 33
584 #define CF_OP_CF_END 34
585 #define CF_OP_LDS_DEALLOC 35
586 #define CF_OP_PUSH_WQM 36
587 #define CF_OP_POP_WQM 37
588 #define CF_OP_ELSE_WQM 38
589 #define CF_OP_JUMP_ANY 39
590 #define CF_OP_REACTIVATE 40
591 #define CF_OP_REACTIVATE_WQM 41
592 #define CF_OP_INTERRUPT 42
593 #define CF_OP_INTERRUPT_AND_SLEEP 43
594 #define CF_OP_SET_PRIORITY 44
595 #define CF_OP_MEM_STREAM0_BUF0 45
596 #define CF_OP_MEM_STREAM0_BUF1 46
597 #define CF_OP_MEM_STREAM0_BUF2 47
598 #define CF_OP_MEM_STREAM0_BUF3 48
599 #define CF_OP_MEM_STREAM1_BUF0 49
600 #define CF_OP_MEM_STREAM1_BUF1 50
601 #define CF_OP_MEM_STREAM1_BUF2 51
602 #define CF_OP_MEM_STREAM1_BUF3 52
603 #define CF_OP_MEM_STREAM2_BUF0 53
604 #define CF_OP_MEM_STREAM2_BUF1 54
605 #define CF_OP_MEM_STREAM2_BUF2 55
606 #define CF_OP_MEM_STREAM2_BUF3 56
607 #define CF_OP_MEM_STREAM3_BUF0 57
608 #define CF_OP_MEM_STREAM3_BUF1 58
609 #define CF_OP_MEM_STREAM3_BUF2 59
610 #define CF_OP_MEM_STREAM3_BUF3 60
611 #define CF_OP_MEM_STREAM0 61
612 #define CF_OP_MEM_STREAM1 62
613 #define CF_OP_MEM_STREAM2 63
614 #define CF_OP_MEM_STREAM3 64
615 #define CF_OP_MEM_SCRATCH 65
616 #define CF_OP_MEM_REDUCT 66
617 #define CF_OP_MEM_RING 67
618 #define CF_OP_EXPORT 68
619 #define CF_OP_EXPORT_DONE 69
620 #define CF_OP_MEM_EXPORT 70
621 #define CF_OP_MEM_RAT 71
622 #define CF_OP_MEM_RAT_NOCACHE 72
623 #define CF_OP_MEM_RING1 73
624 #define CF_OP_MEM_RING2 74
625 #define CF_OP_MEM_RING3 75
626 #define CF_OP_MEM_MEM_COMBINED 76
627 #define CF_OP_MEM_RAT_COMBINED_NOCACHE 77
628 #define CF_OP_MEM_RAT_COMBINED 78
629 #define CF_OP_EXPORT_DONE_END 79
630 #define CF_OP_ALU 80
631 #define CF_OP_ALU_PUSH_BEFORE 81
632 #define CF_OP_ALU_POP_AFTER 82
633 #define CF_OP_ALU_POP2_AFTER 83
634 #define CF_OP_ALU_EXT 84
635 #define CF_OP_ALU_CONTINUE 85
636 #define CF_OP_ALU_BREAK 86
637 #define CF_OP_ALU_VALID_PIXEL_MODE 87
638 #define CF_OP_ALU_ELSE_AFTER 88
639
640 /* CF_NATIVE means that r600_bytecode_cf contains pre-encoded native data */
641 #define CF_NATIVE 89
642
643 enum r600_chip_class {
644 ISA_CC_R600,
645 ISA_CC_R700,
646 ISA_CC_EVERGREEN,
647 ISA_CC_CAYMAN
648 };
649
650 struct r600_isa {
651 enum r600_chip_class hw_class;
652
653 /* these arrays provide reverse mapping - opcode => table_index,
654 * typically we don't need such lookup, unless we are decoding the native
655 * bytecode (e.g. when reading the bytestream from llvm backend) */
656 unsigned *alu_op2_map;
657 unsigned *alu_op3_map;
658 unsigned *fetch_map;
659 unsigned *cf_map;
660 };
661
662 struct r600_context;
663
664 int r600_isa_init(enum amd_gfx_level gfx_level, struct r600_isa *isa);
665 int r600_isa_destroy(struct r600_isa *isa);
666
667 extern const struct alu_op_info r600_alu_op_table[];
668
669 unsigned
670 r600_alu_op_table_size(void);
671
672 const struct alu_op_info *
673 r600_isa_alu(unsigned op);
674
675 const struct fetch_op_info *
676 r600_isa_fetch(unsigned op);
677
678 const struct cf_op_info *
679 r600_isa_cf(unsigned op);
680
681 static inline unsigned
r600_isa_alu_opcode(enum r600_chip_class gfx_level,unsigned op)682 r600_isa_alu_opcode(enum r600_chip_class gfx_level, unsigned op) {
683 int opc = r600_isa_alu(op)->opcode[gfx_level >> 1];
684 assert(opc != -1);
685 return opc;
686 }
687
688 static inline unsigned
r600_isa_alu_slots(enum r600_chip_class gfx_level,unsigned op)689 r600_isa_alu_slots(enum r600_chip_class gfx_level, unsigned op) {
690 unsigned slots = r600_isa_alu(op)->slots[gfx_level];
691 assert(slots != 0);
692 return slots;
693 }
694
695 static inline unsigned
r600_isa_fetch_opcode(enum r600_chip_class gfx_level,unsigned op)696 r600_isa_fetch_opcode(enum r600_chip_class gfx_level, unsigned op) {
697 int opc = r600_isa_fetch(op)->opcode[gfx_level];
698 assert(opc != -1);
699 return opc;
700 }
701
702 static inline unsigned
r600_isa_cf_opcode(enum r600_chip_class gfx_level,unsigned op)703 r600_isa_cf_opcode(enum r600_chip_class gfx_level, unsigned op) {
704 int opc = r600_isa_cf(op)->opcode[gfx_level];
705 assert(opc != -1);
706 return opc;
707 }
708
709 static inline unsigned
r600_isa_alu_by_opcode(struct r600_isa * isa,unsigned opcode,unsigned is_op3)710 r600_isa_alu_by_opcode(struct r600_isa* isa, unsigned opcode, unsigned is_op3) {
711 unsigned op;
712 if (is_op3) {
713 assert(isa->alu_op3_map);
714 op = isa->alu_op3_map[opcode];
715 } else {
716 assert(isa->alu_op2_map);
717 op = isa->alu_op2_map[opcode];
718 }
719 assert(op);
720 return op - 1;
721 }
722
723 static inline unsigned
r600_isa_fetch_by_opcode(struct r600_isa * isa,unsigned opcode)724 r600_isa_fetch_by_opcode(struct r600_isa* isa, unsigned opcode) {
725 unsigned op;
726 assert(isa->fetch_map);
727 op = isa->fetch_map[opcode];
728 assert(op);
729 return op - 1;
730 }
731
732 static inline unsigned
r600_isa_cf_by_opcode(struct r600_isa * isa,unsigned opcode,unsigned is_alu)733 r600_isa_cf_by_opcode(struct r600_isa* isa, unsigned opcode, unsigned is_alu) {
734 unsigned op;
735 assert(isa->cf_map);
736 /* using offset for CF_ALU_xxx opcodes because they overlap with other
737 * CF opcodes (they use different encoding in hw) */
738 op = isa->cf_map[is_alu ? opcode + 0x80 : opcode];
739 assert(op);
740 return op - 1;
741 }
742
743 #ifdef __cplusplus
744 } /* extern "C" */
745 #endif
746
747 #endif /* R600_ISA_H_ */
748